Do not recompile identical OFG in same graph

[ricardoV94]

Draft: as it includes commits from #2056

Repeated OpFromGraph make a lot of sense for deep machine learning graphs. This PR makes more use of this functionality.

In the JIT backends: don't redefine / rewrite the same fgraph.

Numba compile is usually slower than our rewrites, and our caching mechanism already "deduplicates" the functions numba sees. So this is only a %speedup never much absolute. (Big speedup with cache=False, but we are not optimizing for that if it was our goal)

---

Also when dprint don't show identical copies:

import pytensor
import pytensor.tensor as pt

x = pt.vector("x")
relu_op = pytensor.OpFromGraph([x], [pt.maximum(x, 0)], inline=False, name="Relu")

x = pt.vector("x")
a = relu_op(x)
b = relu_op(a)
c = relu_op(b)
out = pt.exp(c)

out.dprint()

Before:

Exp [id A]
 └─ Relu{inline=False} [id B]
    └─ Relu{inline=False} [id C]
       └─ Relu{inline=False} [id D]
          └─ x [id E]

Inner graphs:

Relu{inline=False} [id B]
 ← Maximum [id F]
    ├─ i0 [id G]
    └─ ExpandDims{axis=0} [id H]
       └─ 0 [id I]

Relu{inline=False} [id C]
 ← Maximum [id F]
    └─ ···

Relu{inline=False} [id D]
 ← Maximum [id F]
    └─ ···

After:

Exp [id A]
 └─ Relu{inline=False} [id B]
    └─ Relu{inline=False} [id C]
       └─ Relu{inline=False} [id D]
          └─ x [id E]

Inner graphs:

Relu{inline=False} [id B, C, D]
 ← Maximum [id F]
    ├─ i0 [id G]
    └─ ExpandDims{axis=0} [id H]
       └─ 0 [id I]